techniques like momentum to help the optimizer escape the local minimas. Sharpness-Aware Minimization (SAM)22 is one such technique. It suggests that steep valleys in the objective function might just be the local minima (right). Source: Forret et al. SAM encourages the optimizer to find a minima where the neighborhood of that minima has low loss too, by using the SAM loss. If we denote the weights of a model model by , and its loss function on a training dataset as . Then the SAM objective function is defined as: 23 https://en.wikipedia.org/wiki/Occam%27s_razor 22 Foret, Pierre, et al. "Sharpness-Aware Minimization